block/iscsi.c

   1 /*
   2  * QEMU Block driver for iSCSI images
   3  *
   4  * Copyright (c) 2010-2011 Ronnie Sahlberg <ronniesahlberg@gmail.com>
   5  * Copyright (c) 2012-2014 Peter Lieven <pl@kamp.de>
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a copy
   8  * of this software and associated documentation files (the "Software"), to deal
   9  * in the Software without restriction, including without limitation the rights
  10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  11  * copies of the Software, and to permit persons to whom the Software is
  12  * furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included in
  15  * all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  23  * THE SOFTWARE.
  24  */
  25
  26 #include "config-host.h"
  27
  28 #include <poll.h>
  29 #include <math.h>
  30 #include <arpa/inet.h>
  31 #include "qemu-common.h"
  32 #include "qemu/config-file.h"
  33 #include "qemu/error-report.h"
  34 #include "qemu/bitops.h"
  35 #include "qemu/bitmap.h"
  36 #include "block/block_int.h"
  37 #include "trace.h"
  38 #include "block/scsi.h"
  39 #include "qemu/iov.h"
  40 #include "sysemu/sysemu.h"
  41 #include "qmp-commands.h"
  42
  43 #include <iscsi/iscsi.h>
  44 #include <iscsi/scsi-lowlevel.h>
  45
  46 #ifdef __linux__
  47 #include <scsi/sg.h>
  48 #include <block/scsi.h>
  49 #endif
  50
  51 typedef struct IscsiLun {
  52     struct iscsi_context *iscsi;
  53     AioContext *aio_context;
  54     int lun;
  55     enum scsi_inquiry_peripheral_device_type type;
  56     int block_size;
  57     uint64_t num_blocks;
  58     int events;
  59     QEMUTimer *nop_timer;
  60     uint8_t lbpme;
  61     uint8_t lbprz;
  62     uint8_t has_write_same;
  63     struct scsi_inquiry_logical_block_provisioning lbp;
  64     struct scsi_inquiry_block_limits bl;
  65     unsigned char *zeroblock;
  66     unsigned long *allocationmap;
  67     int cluster_sectors;
  68     bool use_16_for_rw;
  69 } IscsiLun;
  70
  71 typedef struct IscsiTask {
  72     int status;
  73     int complete;
  74     int retries;
  75     int do_retry;
  76     struct scsi_task *task;
  77     Coroutine *co;
  78     QEMUBH *bh;
  79     IscsiLun *iscsilun;
  80     QEMUTimer retry_timer;
  81 } IscsiTask;
  82
  83 typedef struct IscsiAIOCB {
  84     BlockDriverAIOCB common;
  85     QEMUIOVector *qiov;
  86     QEMUBH *bh;
  87     IscsiLun *iscsilun;
  88     struct scsi_task *task;
  89     uint8_t *buf;
  90     int status;
  91     int64_t sector_num;
  92     int nb_sectors;
  93 #ifdef __linux__
  94     sg_io_hdr_t *ioh;
  95 #endif
  96 } IscsiAIOCB;
  97
  98 #define NOP_INTERVAL 5000
  99 #define MAX_NOP_FAILURES 3
 100 #define ISCSI_CMD_RETRIES ARRAY_SIZE(iscsi_retry_times)
 101 static const unsigned iscsi_retry_times[] = {8, 32, 128, 512, 2048};
 102
 103 /* this threshold is a trade-off knob to choose between
 104  * the potential additional overhead of an extra GET_LBA_STATUS request
 105  * vs. unnecessarily reading a lot of zero sectors over the wire.
 106  * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
 107  * sectors we check the allocation status of the area covered by the
 108  * request first if the allocationmap indicates that the area might be
 109  * unallocated. */
 110 #define ISCSI_CHECKALLOC_THRES 64
 111
 112 static void
 113 iscsi_bh_cb(void *p)
 114 {
 115     IscsiAIOCB *acb = p;
 116
 117     qemu_bh_delete(acb->bh);
 118
 119     g_free(acb->buf);
 120     acb->buf = NULL;
 121
 122     acb->common.cb(acb->common.opaque, acb->status);
 123
 124     if (acb->task != NULL) {
 125         scsi_free_scsi_task(acb->task);
 126         acb->task = NULL;
 127     }
 128
 129     qemu_aio_unref(acb);
 130 }
 131
 132 static void
 133 iscsi_schedule_bh(IscsiAIOCB *acb)
 134 {
 135     if (acb->bh) {
 136         return;
 137     }
 138     acb->bh = aio_bh_new(acb->iscsilun->aio_context, iscsi_bh_cb, acb);
 139     qemu_bh_schedule(acb->bh);
 140 }
 141
 142 static void iscsi_co_generic_bh_cb(void *opaque)
 143 {
 144     struct IscsiTask *iTask = opaque;
 145     iTask->complete = 1;
 146     qemu_bh_delete(iTask->bh);
 147     qemu_coroutine_enter(iTask->co, NULL);
 148 }
 149
 150 static void iscsi_retry_timer_expired(void *opaque)
 151 {
 152     struct IscsiTask *iTask = opaque;
 153     iTask->complete = 1;
 154     if (iTask->co) {
 155         qemu_coroutine_enter(iTask->co, NULL);
 156     }
 157 }
 158
 159 static inline unsigned exp_random(double mean)
 160 {
 161     return -mean * log((double)rand() / RAND_MAX);
 162 }
 163
 164 static void
 165 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
 166                         void *command_data, void *opaque)
 167 {
 168     struct IscsiTask *iTask = opaque;
 169     struct scsi_task *task = command_data;
 170
 171     iTask->status = status;
 172     iTask->do_retry = 0;
 173     iTask->task = task;
 174
 175     if (status != SCSI_STATUS_GOOD) {
 176         if (iTask->retries++ < ISCSI_CMD_RETRIES) {
 177             if (status == SCSI_STATUS_CHECK_CONDITION
 178                 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
 179                 error_report("iSCSI CheckCondition: %s",
 180                              iscsi_get_error(iscsi));
 181                 iTask->do_retry = 1;
 182                 goto out;
 183             }
 184             if (status == SCSI_STATUS_BUSY) {
 185                 unsigned retry_time =
 186                     exp_random(iscsi_retry_times[iTask->retries - 1]);
 187                 error_report("iSCSI Busy (retry #%u in %u ms): %s",
 188                              iTask->retries, retry_time,
 189                              iscsi_get_error(iscsi));
 190                 aio_timer_init(iTask->iscsilun->aio_context,
 191                                &iTask->retry_timer, QEMU_CLOCK_REALTIME,
 192                                SCALE_MS, iscsi_retry_timer_expired, iTask);
 193                 timer_mod(&iTask->retry_timer,
 194                           qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + retry_time);
 195                 iTask->do_retry = 1;
 196                 return;
 197             }
 198         }
 199         error_report("iSCSI Failure: %s", iscsi_get_error(iscsi));
 200     }
 201
 202 out:
 203     if (iTask->co) {
 204         iTask->bh = aio_bh_new(iTask->iscsilun->aio_context,
 205                                iscsi_co_generic_bh_cb, iTask);
 206         qemu_bh_schedule(iTask->bh);
 207     } else {
 208         iTask->complete = 1;
 209     }
 210 }
 211
 212 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
 213 {
 214     *iTask = (struct IscsiTask) {
 215         .co         = qemu_coroutine_self(),
 216         .iscsilun   = iscsilun,
 217     };
 218 }
 219
 220 static void
 221 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
 222                     void *private_data)
 223 {
 224     IscsiAIOCB *acb = private_data;
 225
 226     acb->status = -ECANCELED;
 227     iscsi_schedule_bh(acb);
 228 }
 229
 230 static void
 231 iscsi_aio_cancel(BlockDriverAIOCB *blockacb)
 232 {
 233     IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
 234     IscsiLun *iscsilun = acb->iscsilun;
 235
 236     if (acb->status != -EINPROGRESS) {
 237         return;
 238     }
 239
 240     /* send a task mgmt call to the target to cancel the task on the target */
 241     iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
 242                                      iscsi_abort_task_cb, acb);
 243
 244 }
 245
 246 static const AIOCBInfo iscsi_aiocb_info = {
 247     .aiocb_size         = sizeof(IscsiAIOCB),
 248     .cancel_async       = iscsi_aio_cancel,
 249 };
 250
 251
 252 static void iscsi_process_read(void *arg);
 253 static void iscsi_process_write(void *arg);
 254
 255 static void
 256 iscsi_set_events(IscsiLun *iscsilun)
 257 {
 258     struct iscsi_context *iscsi = iscsilun->iscsi;
 259     int ev;
 260
 261     /* We always register a read handler.  */
 262     ev = POLLIN;
 263     ev |= iscsi_which_events(iscsi);
 264     if (ev != iscsilun->events) {
 265         aio_set_fd_handler(iscsilun->aio_context,
 266                            iscsi_get_fd(iscsi),
 267                            iscsi_process_read,
 268                            (ev & POLLOUT) ? iscsi_process_write : NULL,
 269                            iscsilun);
 270
 271     }
 272
 273     iscsilun->events = ev;
 274 }
 275
 276 static void
 277 iscsi_process_read(void *arg)
 278 {
 279     IscsiLun *iscsilun = arg;
 280     struct iscsi_context *iscsi = iscsilun->iscsi;
 281
 282     iscsi_service(iscsi, POLLIN);
 283     iscsi_set_events(iscsilun);
 284 }
 285
 286 static void
 287 iscsi_process_write(void *arg)
 288 {
 289     IscsiLun *iscsilun = arg;
 290     struct iscsi_context *iscsi = iscsilun->iscsi;
 291
 292     iscsi_service(iscsi, POLLOUT);
 293     iscsi_set_events(iscsilun);
 294 }
 295
 296 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
 297 {
 298     return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
 299 }
 300
 301 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
 302 {
 303     return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
 304 }
 305
 306 static bool is_request_lun_aligned(int64_t sector_num, int nb_sectors,
 307                                       IscsiLun *iscsilun)
 308 {
 309     if ((sector_num * BDRV_SECTOR_SIZE) % iscsilun->block_size ||
 310         (nb_sectors * BDRV_SECTOR_SIZE) % iscsilun->block_size) {
 311             error_report("iSCSI misaligned request: "
 312                          "iscsilun->block_size %u, sector_num %" PRIi64
 313                          ", nb_sectors %d",
 314                          iscsilun->block_size, sector_num, nb_sectors);
 315             return 0;
 316     }
 317     return 1;
 318 }
 319
 320 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 321                                     int nb_sectors)
 322 {
 323     if (iscsilun->allocationmap == NULL) {
 324         return;
 325     }
 326     bitmap_set(iscsilun->allocationmap,
 327                sector_num / iscsilun->cluster_sectors,
 328                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 329 }
 330
 331 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 332                                       int nb_sectors)
 333 {
 334     int64_t cluster_num, nb_clusters;
 335     if (iscsilun->allocationmap == NULL) {
 336         return;
 337     }
 338     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 339     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 340                   - cluster_num;
 341     if (nb_clusters > 0) {
 342         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 343     }
 344 }
 345
 346 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 347                                         int64_t sector_num, int nb_sectors,
 348                                         QEMUIOVector *iov)
 349 {
 350     IscsiLun *iscsilun = bs->opaque;
 351     struct IscsiTask iTask;
 352     uint64_t lba;
 353     uint32_t num_sectors;
 354
 355     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 356         return -EINVAL;
 357     }
 358
 359     lba = sector_qemu2lun(sector_num, iscsilun);
 360     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 361     iscsi_co_init_iscsitask(iscsilun, &iTask);
 362 retry:
 363     if (iscsilun->use_16_for_rw) {
 364         iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 365                                         NULL, num_sectors * iscsilun->block_size,
 366                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 367                                         iscsi_co_generic_cb, &iTask);
 368     } else {
 369         iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
 370                                         NULL, num_sectors * iscsilun->block_size,
 371                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 372                                         iscsi_co_generic_cb, &iTask);
 373     }
 374     if (iTask.task == NULL) {
 375         return -ENOMEM;
 376     }
 377     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 378                           iov->niov);
 379     while (!iTask.complete) {
 380         iscsi_set_events(iscsilun);
 381         qemu_coroutine_yield();
 382     }
 383
 384     if (iTask.task != NULL) {
 385         scsi_free_scsi_task(iTask.task);
 386         iTask.task = NULL;
 387     }
 388
 389     if (iTask.do_retry) {
 390         iTask.complete = 0;
 391         goto retry;
 392     }
 393
 394     if (iTask.status != SCSI_STATUS_GOOD) {
 395         return -EIO;
 396     }
 397
 398     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 399
 400     return 0;
 401 }
 402
 403
 404 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 405                                              int64_t sector_num, int nb_sectors)
 406 {
 407     unsigned long size;
 408     if (iscsilun->allocationmap == NULL) {
 409         return true;
 410     }
 411     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 412     return !(find_next_bit(iscsilun->allocationmap, size,
 413                            sector_num / iscsilun->cluster_sectors) == size);
 414 }
 415
 416 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 417                                                   int64_t sector_num,
 418                                                   int nb_sectors, int *pnum)
 419 {
 420     IscsiLun *iscsilun = bs->opaque;
 421     struct scsi_get_lba_status *lbas = NULL;
 422     struct scsi_lba_status_descriptor *lbasd = NULL;
 423     struct IscsiTask iTask;
 424     int64_t ret;
 425
 426     iscsi_co_init_iscsitask(iscsilun, &iTask);
 427
 428     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 429         ret = -EINVAL;
 430         goto out;
 431     }
 432
 433     /* default to all sectors allocated */
 434     ret = BDRV_BLOCK_DATA;
 435     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 436     *pnum = nb_sectors;
 437
 438     /* LUN does not support logical block provisioning */
 439     if (iscsilun->lbpme == 0) {
 440         goto out;
 441     }
 442
 443 retry:
 444     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 445                                   sector_qemu2lun(sector_num, iscsilun),
 446                                   8 + 16, iscsi_co_generic_cb,
 447                                   &iTask) == NULL) {
 448         ret = -ENOMEM;
 449         goto out;
 450     }
 451
 452     while (!iTask.complete) {
 453         iscsi_set_events(iscsilun);
 454         qemu_coroutine_yield();
 455     }
 456
 457     if (iTask.do_retry) {
 458         if (iTask.task != NULL) {
 459             scsi_free_scsi_task(iTask.task);
 460             iTask.task = NULL;
 461         }
 462         iTask.complete = 0;
 463         goto retry;
 464     }
 465
 466     if (iTask.status != SCSI_STATUS_GOOD) {
 467         /* in case the get_lba_status_callout fails (i.e.
 468          * because the device is busy or the cmd is not
 469          * supported) we pretend all blocks are allocated
 470          * for backwards compatibility */
 471         goto out;
 472     }
 473
 474     lbas = scsi_datain_unmarshall(iTask.task);
 475     if (lbas == NULL) {
 476         ret = -EIO;
 477         goto out;
 478     }
 479
 480     lbasd = &lbas->descriptors[0];
 481
 482     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 483         ret = -EIO;
 484         goto out;
 485     }
 486
 487     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 488
 489     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 490         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 491         ret &= ~BDRV_BLOCK_DATA;
 492         if (iscsilun->lbprz) {
 493             ret |= BDRV_BLOCK_ZERO;
 494         }
 495     }
 496
 497     if (ret & BDRV_BLOCK_ZERO) {
 498         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 499     } else {
 500         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 501     }
 502
 503     if (*pnum > nb_sectors) {
 504         *pnum = nb_sectors;
 505     }
 506 out:
 507     if (iTask.task != NULL) {
 508         scsi_free_scsi_task(iTask.task);
 509     }
 510     return ret;
 511 }
 512
 513 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 514                                        int64_t sector_num, int nb_sectors,
 515                                        QEMUIOVector *iov)
 516 {
 517     IscsiLun *iscsilun = bs->opaque;
 518     struct IscsiTask iTask;
 519     uint64_t lba;
 520     uint32_t num_sectors;
 521
 522     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 523         return -EINVAL;
 524     }
 525
 526     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 527         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 528         int64_t ret;
 529         int pnum;
 530         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 531         if (ret < 0) {
 532             return ret;
 533         }
 534         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 535             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 536             return 0;
 537         }
 538     }
 539
 540     lba = sector_qemu2lun(sector_num, iscsilun);
 541     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 542
 543     iscsi_co_init_iscsitask(iscsilun, &iTask);
 544 retry:
 545     if (iscsilun->use_16_for_rw) {
 546         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 547                                        num_sectors * iscsilun->block_size,
 548                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 549                                        iscsi_co_generic_cb, &iTask);
 550     } else {
 551         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 552                                        num_sectors * iscsilun->block_size,
 553                                        iscsilun->block_size,
 554                                        0, 0, 0, 0, 0,
 555                                        iscsi_co_generic_cb, &iTask);
 556     }
 557     if (iTask.task == NULL) {
 558         return -ENOMEM;
 559     }
 560     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 561
 562     while (!iTask.complete) {
 563         iscsi_set_events(iscsilun);
 564         qemu_coroutine_yield();
 565     }
 566
 567     if (iTask.task != NULL) {
 568         scsi_free_scsi_task(iTask.task);
 569         iTask.task = NULL;
 570     }
 571
 572     if (iTask.do_retry) {
 573         iTask.complete = 0;
 574         goto retry;
 575     }
 576
 577     if (iTask.status != SCSI_STATUS_GOOD) {
 578         return -EIO;
 579     }
 580
 581     return 0;
 582 }
 583
 584 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 585 {
 586     IscsiLun *iscsilun = bs->opaque;
 587     struct IscsiTask iTask;
 588
 589     if (bs->sg) {
 590         return 0;
 591     }
 592
 593     iscsi_co_init_iscsitask(iscsilun, &iTask);
 594
 595 retry:
 596     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 597                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 598         return -ENOMEM;
 599     }
 600
 601     while (!iTask.complete) {
 602         iscsi_set_events(iscsilun);
 603         qemu_coroutine_yield();
 604     }
 605
 606     if (iTask.task != NULL) {
 607         scsi_free_scsi_task(iTask.task);
 608         iTask.task = NULL;
 609     }
 610
 611     if (iTask.do_retry) {
 612         iTask.complete = 0;
 613         goto retry;
 614     }
 615
 616     if (iTask.status != SCSI_STATUS_GOOD) {
 617         return -EIO;
 618     }
 619
 620     return 0;
 621 }
 622
 623 #ifdef __linux__
 624 static void
 625 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 626                      void *command_data, void *opaque)
 627 {
 628     IscsiAIOCB *acb = opaque;
 629
 630     g_free(acb->buf);
 631     acb->buf = NULL;
 632
 633     acb->status = 0;
 634     if (status < 0) {
 635         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 636                      iscsi_get_error(iscsi));
 637         acb->status = -EIO;
 638     }
 639
 640     acb->ioh->driver_status = 0;
 641     acb->ioh->host_status   = 0;
 642     acb->ioh->resid         = 0;
 643
 644 #define SG_ERR_DRIVER_SENSE    0x08
 645
 646     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 647         int ss;
 648
 649         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 650
 651         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 652         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 653              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 654         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 655     }
 656
 657     iscsi_schedule_bh(acb);
 658 }
 659
 660 static BlockDriverAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 661         unsigned long int req, void *buf,
 662         BlockDriverCompletionFunc *cb, void *opaque)
 663 {
 664     IscsiLun *iscsilun = bs->opaque;
 665     struct iscsi_context *iscsi = iscsilun->iscsi;
 666     struct iscsi_data data;
 667     IscsiAIOCB *acb;
 668
 669     assert(req == SG_IO);
 670
 671     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 672
 673     acb->iscsilun = iscsilun;
 674     acb->bh          = NULL;
 675     acb->status      = -EINPROGRESS;
 676     acb->buf         = NULL;
 677     acb->ioh         = buf;
 678
 679     acb->task = malloc(sizeof(struct scsi_task));
 680     if (acb->task == NULL) {
 681         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 682                      iscsi_get_error(iscsi));
 683         qemu_aio_unref(acb);
 684         return NULL;
 685     }
 686     memset(acb->task, 0, sizeof(struct scsi_task));
 687
 688     switch (acb->ioh->dxfer_direction) {
 689     case SG_DXFER_TO_DEV:
 690         acb->task->xfer_dir = SCSI_XFER_WRITE;
 691         break;
 692     case SG_DXFER_FROM_DEV:
 693         acb->task->xfer_dir = SCSI_XFER_READ;
 694         break;
 695     default:
 696         acb->task->xfer_dir = SCSI_XFER_NONE;
 697         break;
 698     }
 699
 700     acb->task->cdb_size = acb->ioh->cmd_len;
 701     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 702     acb->task->expxferlen = acb->ioh->dxfer_len;
 703
 704     data.size = 0;
 705     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 706         if (acb->ioh->iovec_count == 0) {
 707             data.data = acb->ioh->dxferp;
 708             data.size = acb->ioh->dxfer_len;
 709         } else {
 710             scsi_task_set_iov_out(acb->task,
 711                                  (struct scsi_iovec *) acb->ioh->dxferp,
 712                                  acb->ioh->iovec_count);
 713         }
 714     }
 715
 716     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 717                                  iscsi_aio_ioctl_cb,
 718                                  (data.size > 0) ? &data : NULL,
 719                                  acb) != 0) {
 720         scsi_free_scsi_task(acb->task);
 721         qemu_aio_unref(acb);
 722         return NULL;
 723     }
 724
 725     /* tell libiscsi to read straight into the buffer we got from ioctl */
 726     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 727         if (acb->ioh->iovec_count == 0) {
 728             scsi_task_add_data_in_buffer(acb->task,
 729                                          acb->ioh->dxfer_len,
 730                                          acb->ioh->dxferp);
 731         } else {
 732             scsi_task_set_iov_in(acb->task,
 733                                  (struct scsi_iovec *) acb->ioh->dxferp,
 734                                  acb->ioh->iovec_count);
 735         }
 736     }
 737
 738     iscsi_set_events(iscsilun);
 739
 740     return &acb->common;
 741 }
 742
 743 static void ioctl_cb(void *opaque, int status)
 744 {
 745     int *p_status = opaque;
 746     *p_status = status;
 747 }
 748
 749 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 750 {
 751     IscsiLun *iscsilun = bs->opaque;
 752     int status;
 753
 754     switch (req) {
 755     case SG_GET_VERSION_NUM:
 756         *(int *)buf = 30000;
 757         break;
 758     case SG_GET_SCSI_ID:
 759         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 760         break;
 761     case SG_IO:
 762         status = -EINPROGRESS;
 763         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 764
 765         while (status == -EINPROGRESS) {
 766             aio_poll(iscsilun->aio_context, true);
 767         }
 768
 769         return 0;
 770     default:
 771         return -1;
 772     }
 773     return 0;
 774 }
 775 #endif
 776
 777 static int64_t
 778 iscsi_getlength(BlockDriverState *bs)
 779 {
 780     IscsiLun *iscsilun = bs->opaque;
 781     int64_t len;
 782
 783     len  = iscsilun->num_blocks;
 784     len *= iscsilun->block_size;
 785
 786     return len;
 787 }
 788
 789 static int
 790 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 791                                    int nb_sectors)
 792 {
 793     IscsiLun *iscsilun = bs->opaque;
 794     struct IscsiTask iTask;
 795     struct unmap_list list;
 796
 797     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 798         return -EINVAL;
 799     }
 800
 801     if (!iscsilun->lbp.lbpu) {
 802         /* UNMAP is not supported by the target */
 803         return 0;
 804     }
 805
 806     list.lba = sector_qemu2lun(sector_num, iscsilun);
 807     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 808
 809     iscsi_co_init_iscsitask(iscsilun, &iTask);
 810 retry:
 811     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 812                      iscsi_co_generic_cb, &iTask) == NULL) {
 813         return -ENOMEM;
 814     }
 815
 816     while (!iTask.complete) {
 817         iscsi_set_events(iscsilun);
 818         qemu_coroutine_yield();
 819     }
 820
 821     if (iTask.task != NULL) {
 822         scsi_free_scsi_task(iTask.task);
 823         iTask.task = NULL;
 824     }
 825
 826     if (iTask.do_retry) {
 827         iTask.complete = 0;
 828         goto retry;
 829     }
 830
 831     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 832         /* the target might fail with a check condition if it
 833            is not happy with the alignment of the UNMAP request
 834            we silently fail in this case */
 835         return 0;
 836     }
 837
 838     if (iTask.status != SCSI_STATUS_GOOD) {
 839         return -EIO;
 840     }
 841
 842     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 843
 844     return 0;
 845 }
 846
 847 static int
 848 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 849                                    int nb_sectors, BdrvRequestFlags flags)
 850 {
 851     IscsiLun *iscsilun = bs->opaque;
 852     struct IscsiTask iTask;
 853     uint64_t lba;
 854     uint32_t nb_blocks;
 855     bool use_16_for_ws = iscsilun->use_16_for_rw;
 856
 857     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 858         return -EINVAL;
 859     }
 860
 861     if (flags & BDRV_REQ_MAY_UNMAP) {
 862         if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
 863             /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
 864             use_16_for_ws = true;
 865         }
 866         if (use_16_for_ws && !iscsilun->lbp.lbpws) {
 867             /* WRITESAME16 with UNMAP is not supported by the target,
 868              * fall back and try WRITESAME10/16 without UNMAP */
 869             flags &= ~BDRV_REQ_MAY_UNMAP;
 870             use_16_for_ws = iscsilun->use_16_for_rw;
 871         }
 872     }
 873
 874     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 875         /* WRITESAME without UNMAP is not supported by the target */
 876         return -ENOTSUP;
 877     }
 878
 879     lba = sector_qemu2lun(sector_num, iscsilun);
 880     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 881
 882     if (iscsilun->zeroblock == NULL) {
 883         iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
 884         if (iscsilun->zeroblock == NULL) {
 885             return -ENOMEM;
 886         }
 887     }
 888
 889     iscsi_co_init_iscsitask(iscsilun, &iTask);
 890 retry:
 891     if (use_16_for_ws) {
 892         iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 893                                             iscsilun->zeroblock, iscsilun->block_size,
 894                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 895                                             0, 0, iscsi_co_generic_cb, &iTask);
 896     } else {
 897         iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
 898                                             iscsilun->zeroblock, iscsilun->block_size,
 899                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 900                                             0, 0, iscsi_co_generic_cb, &iTask);
 901     }
 902     if (iTask.task == NULL) {
 903         return -ENOMEM;
 904     }
 905
 906     while (!iTask.complete) {
 907         iscsi_set_events(iscsilun);
 908         qemu_coroutine_yield();
 909     }
 910
 911     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 912         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 913         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 914          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 915         /* WRITE SAME is not supported by the target */
 916         iscsilun->has_write_same = false;
 917         scsi_free_scsi_task(iTask.task);
 918         return -ENOTSUP;
 919     }
 920
 921     if (iTask.task != NULL) {
 922         scsi_free_scsi_task(iTask.task);
 923         iTask.task = NULL;
 924     }
 925
 926     if (iTask.do_retry) {
 927         iTask.complete = 0;
 928         goto retry;
 929     }
 930
 931     if (iTask.status != SCSI_STATUS_GOOD) {
 932         return -EIO;
 933     }
 934
 935     if (flags & BDRV_REQ_MAY_UNMAP) {
 936         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 937     } else {
 938         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 939     }
 940
 941     return 0;
 942 }
 943
 944 static void parse_chap(struct iscsi_context *iscsi, const char *target,
 945                        Error **errp)
 946 {
 947     QemuOptsList *list;
 948     QemuOpts *opts;
 949     const char *user = NULL;
 950     const char *password = NULL;
 951
 952     list = qemu_find_opts("iscsi");
 953     if (!list) {
 954         return;
 955     }
 956
 957     opts = qemu_opts_find(list, target);
 958     if (opts == NULL) {
 959         opts = QTAILQ_FIRST(&list->head);
 960         if (!opts) {
 961             return;
 962         }
 963     }
 964
 965     user = qemu_opt_get(opts, "user");
 966     if (!user) {
 967         return;
 968     }
 969
 970     password = qemu_opt_get(opts, "password");
 971     if (!password) {
 972         error_setg(errp, "CHAP username specified but no password was given");
 973         return;
 974     }
 975
 976     if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
 977         error_setg(errp, "Failed to set initiator username and password");
 978     }
 979 }
 980
 981 static void parse_header_digest(struct iscsi_context *iscsi, const char *target,
 982                                 Error **errp)
 983 {
 984     QemuOptsList *list;
 985     QemuOpts *opts;
 986     const char *digest = NULL;
 987
 988     list = qemu_find_opts("iscsi");
 989     if (!list) {
 990         return;
 991     }
 992
 993     opts = qemu_opts_find(list, target);
 994     if (opts == NULL) {
 995         opts = QTAILQ_FIRST(&list->head);
 996         if (!opts) {
 997             return;
 998         }
 999     }
1000
1001     digest = qemu_opt_get(opts, "header-digest");
1002     if (!digest) {
1003         return;
1004     }
1005
1006     if (!strcmp(digest, "CRC32C")) {
1007         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1008     } else if (!strcmp(digest, "NONE")) {
1009         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1010     } else if (!strcmp(digest, "CRC32C-NONE")) {
1011         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1012     } else if (!strcmp(digest, "NONE-CRC32C")) {
1013         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1014     } else {
1015         error_setg(errp, "Invalid header-digest setting : %s", digest);
1016     }
1017 }
1018
1019 static char *parse_initiator_name(const char *target)
1020 {
1021     QemuOptsList *list;
1022     QemuOpts *opts;
1023     const char *name;
1024     char *iscsi_name;
1025     UuidInfo *uuid_info;
1026
1027     list = qemu_find_opts("iscsi");
1028     if (list) {
1029         opts = qemu_opts_find(list, target);
1030         if (!opts) {
1031             opts = QTAILQ_FIRST(&list->head);
1032         }
1033         if (opts) {
1034             name = qemu_opt_get(opts, "initiator-name");
1035             if (name) {
1036                 return g_strdup(name);
1037             }
1038         }
1039     }
1040
1041     uuid_info = qmp_query_uuid(NULL);
1042     if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1043         name = qemu_get_vm_name();
1044     } else {
1045         name = uuid_info->UUID;
1046     }
1047     iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1048                                  name ? ":" : "", name ? name : "");
1049     qapi_free_UuidInfo(uuid_info);
1050     return iscsi_name;
1051 }
1052
1053 static void iscsi_nop_timed_event(void *opaque)
1054 {
1055     IscsiLun *iscsilun = opaque;
1056
1057     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1058         error_report("iSCSI: NOP timeout. Reconnecting...");
1059         iscsi_reconnect(iscsilun->iscsi);
1060     }
1061
1062     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1063         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1064         return;
1065     }
1066
1067     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1068     iscsi_set_events(iscsilun);
1069 }
1070
1071 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1072 {
1073     struct scsi_task *task = NULL;
1074     struct scsi_readcapacity10 *rc10 = NULL;
1075     struct scsi_readcapacity16 *rc16 = NULL;
1076     int retries = ISCSI_CMD_RETRIES;
1077
1078     do {
1079         if (task != NULL) {
1080             scsi_free_scsi_task(task);
1081             task = NULL;
1082         }
1083
1084         switch (iscsilun->type) {
1085         case TYPE_DISK:
1086             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1087             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1088                 rc16 = scsi_datain_unmarshall(task);
1089                 if (rc16 == NULL) {
1090                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1091                 } else {
1092                     iscsilun->block_size = rc16->block_length;
1093                     iscsilun->num_blocks = rc16->returned_lba + 1;
1094                     iscsilun->lbpme = rc16->lbpme;
1095                     iscsilun->lbprz = rc16->lbprz;
1096                     iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1097                 }
1098             }
1099             break;
1100         case TYPE_ROM:
1101             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1102             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1103                 rc10 = scsi_datain_unmarshall(task);
1104                 if (rc10 == NULL) {
1105                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1106                 } else {
1107                     iscsilun->block_size = rc10->block_size;
1108                     if (rc10->lba == 0) {
1109                         /* blank disk loaded */
1110                         iscsilun->num_blocks = 0;
1111                     } else {
1112                         iscsilun->num_blocks = rc10->lba + 1;
1113                     }
1114                 }
1115             }
1116             break;
1117         default:
1118             return;
1119         }
1120     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1121              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1122              && retries-- > 0);
1123
1124     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1125         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1126     }
1127     if (task) {
1128         scsi_free_scsi_task(task);
1129     }
1130 }
1131
1132 /* TODO Convert to fine grained options */
1133 static QemuOptsList runtime_opts = {
1134     .name = "iscsi",
1135     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1136     .desc = {
1137         {
1138             .name = "filename",
1139             .type = QEMU_OPT_STRING,
1140             .help = "URL to the iscsi image",
1141         },
1142         { /* end of list */ }
1143     },
1144 };
1145
1146 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1147                                           int evpd, int pc, void **inq, Error **errp)
1148 {
1149     int full_size;
1150     struct scsi_task *task = NULL;
1151     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1152     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1153         goto fail;
1154     }
1155     full_size = scsi_datain_getfullsize(task);
1156     if (full_size > task->datain.size) {
1157         scsi_free_scsi_task(task);
1158
1159         /* we need more data for the full list */
1160         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1161         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1162             goto fail;
1163         }
1164     }
1165
1166     *inq = scsi_datain_unmarshall(task);
1167     if (*inq == NULL) {
1168         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1169         goto fail_with_err;
1170     }
1171
1172     return task;
1173
1174 fail:
1175     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1176                iscsi_get_error(iscsi));
1177 fail_with_err:
1178     if (task != NULL) {
1179         scsi_free_scsi_task(task);
1180     }
1181     return NULL;
1182 }
1183
1184 static void iscsi_detach_aio_context(BlockDriverState *bs)
1185 {
1186     IscsiLun *iscsilun = bs->opaque;
1187
1188     aio_set_fd_handler(iscsilun->aio_context,
1189                        iscsi_get_fd(iscsilun->iscsi),
1190                        NULL, NULL, NULL);
1191     iscsilun->events = 0;
1192
1193     if (iscsilun->nop_timer) {
1194         timer_del(iscsilun->nop_timer);
1195         timer_free(iscsilun->nop_timer);
1196         iscsilun->nop_timer = NULL;
1197     }
1198 }
1199
1200 static void iscsi_attach_aio_context(BlockDriverState *bs,
1201                                      AioContext *new_context)
1202 {
1203     IscsiLun *iscsilun = bs->opaque;
1204
1205     iscsilun->aio_context = new_context;
1206     iscsi_set_events(iscsilun);
1207
1208     /* Set up a timer for sending out iSCSI NOPs */
1209     iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1210                                         QEMU_CLOCK_REALTIME, SCALE_MS,
1211                                         iscsi_nop_timed_event, iscsilun);
1212     timer_mod(iscsilun->nop_timer,
1213               qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1214 }
1215
1216 /*
1217  * We support iscsi url's on the form
1218  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1219  *
1220  * Note: flags are currently not used by iscsi_open.  If this function
1221  * is changed such that flags are used, please examine iscsi_reopen_prepare()
1222  * to see if needs to be changed as well.
1223  */
1224 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1225                       Error **errp)
1226 {
1227     IscsiLun *iscsilun = bs->opaque;
1228     struct iscsi_context *iscsi = NULL;
1229     struct iscsi_url *iscsi_url = NULL;
1230     struct scsi_task *task = NULL;
1231     struct scsi_inquiry_standard *inq = NULL;
1232     struct scsi_inquiry_supported_pages *inq_vpd;
1233     char *initiator_name = NULL;
1234     QemuOpts *opts;
1235     Error *local_err = NULL;
1236     const char *filename;
1237     int i, ret;
1238
1239     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1240         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1241                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1242                    "of 512", BDRV_SECTOR_SIZE);
1243         return -EINVAL;
1244     }
1245
1246     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1247     qemu_opts_absorb_qdict(opts, options, &local_err);
1248     if (local_err) {
1249         error_propagate(errp, local_err);
1250         ret = -EINVAL;
1251         goto out;
1252     }
1253
1254     filename = qemu_opt_get(opts, "filename");
1255
1256     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1257     if (iscsi_url == NULL) {
1258         error_setg(errp, "Failed to parse URL : %s", filename);
1259         ret = -EINVAL;
1260         goto out;
1261     }
1262
1263     memset(iscsilun, 0, sizeof(IscsiLun));
1264
1265     initiator_name = parse_initiator_name(iscsi_url->target);
1266
1267     iscsi = iscsi_create_context(initiator_name);
1268     if (iscsi == NULL) {
1269         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1270         ret = -ENOMEM;
1271         goto out;
1272     }
1273
1274     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1275         error_setg(errp, "iSCSI: Failed to set target name.");
1276         ret = -EINVAL;
1277         goto out;
1278     }
1279
1280     if (iscsi_url->user != NULL) {
1281         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1282                                               iscsi_url->passwd);
1283         if (ret != 0) {
1284             error_setg(errp, "Failed to set initiator username and password");
1285             ret = -EINVAL;
1286             goto out;
1287         }
1288     }
1289
1290     /* check if we got CHAP username/password via the options */
1291     parse_chap(iscsi, iscsi_url->target, &local_err);
1292     if (local_err != NULL) {
1293         error_propagate(errp, local_err);
1294         ret = -EINVAL;
1295         goto out;
1296     }
1297
1298     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1299         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1300         ret = -EINVAL;
1301         goto out;
1302     }
1303
1304     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1305
1306     /* check if we got HEADER_DIGEST via the options */
1307     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1308     if (local_err != NULL) {
1309         error_propagate(errp, local_err);
1310         ret = -EINVAL;
1311         goto out;
1312     }
1313
1314     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1315         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1316             iscsi_get_error(iscsi));
1317         ret = -EINVAL;
1318         goto out;
1319     }
1320
1321     iscsilun->iscsi = iscsi;
1322     iscsilun->aio_context = bdrv_get_aio_context(bs);
1323     iscsilun->lun   = iscsi_url->lun;
1324     iscsilun->has_write_same = true;
1325
1326     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1327                             (void **) &inq, errp);
1328     if (task == NULL) {
1329         ret = -EINVAL;
1330         goto out;
1331     }
1332     iscsilun->type = inq->periperal_device_type;
1333     scsi_free_scsi_task(task);
1334     task = NULL;
1335
1336     iscsi_readcapacity_sync(iscsilun, &local_err);
1337     if (local_err != NULL) {
1338         error_propagate(errp, local_err);
1339         ret = -EINVAL;
1340         goto out;
1341     }
1342     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1343     bs->request_alignment = iscsilun->block_size;
1344
1345     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1346      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1347      * will try to read from the device to guess the image format.
1348      */
1349     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1350         bs->sg = 1;
1351     }
1352
1353     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1354                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1355                             (void **) &inq_vpd, errp);
1356     if (task == NULL) {
1357         ret = -EINVAL;
1358         goto out;
1359     }
1360     for (i = 0; i < inq_vpd->num_pages; i++) {
1361         struct scsi_task *inq_task;
1362         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1363         struct scsi_inquiry_block_limits *inq_bl;
1364         switch (inq_vpd->pages[i]) {
1365         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1366             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1367                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1368                                         (void **) &inq_lbp, errp);
1369             if (inq_task == NULL) {
1370                 ret = -EINVAL;
1371                 goto out;
1372             }
1373             memcpy(&iscsilun->lbp, inq_lbp,
1374                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1375             scsi_free_scsi_task(inq_task);
1376             break;
1377         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1378             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1379                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1380                                     (void **) &inq_bl, errp);
1381             if (inq_task == NULL) {
1382                 ret = -EINVAL;
1383                 goto out;
1384             }
1385             memcpy(&iscsilun->bl, inq_bl,
1386                    sizeof(struct scsi_inquiry_block_limits));
1387             scsi_free_scsi_task(inq_task);
1388             break;
1389         default:
1390             break;
1391         }
1392     }
1393     scsi_free_scsi_task(task);
1394     task = NULL;
1395
1396     iscsi_attach_aio_context(bs, iscsilun->aio_context);
1397
1398     /* Guess the internal cluster (page) size of the iscsi target by the means
1399      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1400      * reasonable size */
1401     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1402         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1403         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1404                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1405         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1406             iscsilun->allocationmap =
1407                 bitmap_new(DIV_ROUND_UP(bs->total_sectors,
1408                                         iscsilun->cluster_sectors));
1409         }
1410     }
1411
1412 out:
1413     qemu_opts_del(opts);
1414     g_free(initiator_name);
1415     if (iscsi_url != NULL) {
1416         iscsi_destroy_url(iscsi_url);
1417     }
1418     if (task != NULL) {
1419         scsi_free_scsi_task(task);
1420     }
1421
1422     if (ret) {
1423         if (iscsi != NULL) {
1424             iscsi_destroy_context(iscsi);
1425         }
1426         memset(iscsilun, 0, sizeof(IscsiLun));
1427     }
1428     return ret;
1429 }
1430
1431 static void iscsi_close(BlockDriverState *bs)
1432 {
1433     IscsiLun *iscsilun = bs->opaque;
1434     struct iscsi_context *iscsi = iscsilun->iscsi;
1435
1436     iscsi_detach_aio_context(bs);
1437     iscsi_destroy_context(iscsi);
1438     g_free(iscsilun->zeroblock);
1439     g_free(iscsilun->allocationmap);
1440     memset(iscsilun, 0, sizeof(IscsiLun));
1441 }
1442
1443 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
1444 {
1445     IscsiLun *iscsilun = bs->opaque;
1446
1447     /* We don't actually refresh here, but just return data queried in
1448      * iscsi_open(): iscsi targets don't change their limits. */
1449     if (iscsilun->lbp.lbpu) {
1450         if (iscsilun->bl.max_unmap < 0xffffffff) {
1451             bs->bl.max_discard = sector_lun2qemu(iscsilun->bl.max_unmap,
1452                                                  iscsilun);
1453         }
1454         bs->bl.discard_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1455                                                    iscsilun);
1456     }
1457
1458     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1459         bs->bl.max_write_zeroes = sector_lun2qemu(iscsilun->bl.max_ws_len,
1460                                                   iscsilun);
1461     }
1462     if (iscsilun->lbp.lbpws) {
1463         bs->bl.write_zeroes_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1464                                                         iscsilun);
1465     }
1466     bs->bl.opt_transfer_length = sector_lun2qemu(iscsilun->bl.opt_xfer_len,
1467                                                  iscsilun);
1468 }
1469
1470 /* Since iscsi_open() ignores bdrv_flags, there is nothing to do here in
1471  * prepare.  Note that this will not re-establish a connection with an iSCSI
1472  * target - it is effectively a NOP.  */
1473 static int iscsi_reopen_prepare(BDRVReopenState *state,
1474                                 BlockReopenQueue *queue, Error **errp)
1475 {
1476     /* NOP */
1477     return 0;
1478 }
1479
1480 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1481 {
1482     IscsiLun *iscsilun = bs->opaque;
1483     Error *local_err = NULL;
1484
1485     if (iscsilun->type != TYPE_DISK) {
1486         return -ENOTSUP;
1487     }
1488
1489     iscsi_readcapacity_sync(iscsilun, &local_err);
1490     if (local_err != NULL) {
1491         error_free(local_err);
1492         return -EIO;
1493     }
1494
1495     if (offset > iscsi_getlength(bs)) {
1496         return -EINVAL;
1497     }
1498
1499     if (iscsilun->allocationmap != NULL) {
1500         g_free(iscsilun->allocationmap);
1501         iscsilun->allocationmap =
1502             bitmap_new(DIV_ROUND_UP(sector_lun2qemu(iscsilun->num_blocks,
1503                                                     iscsilun),
1504                                     iscsilun->cluster_sectors));
1505     }
1506
1507     return 0;
1508 }
1509
1510 static int iscsi_create(const char *filename, QemuOpts *opts, Error **errp)
1511 {
1512     int ret = 0;
1513     int64_t total_size = 0;
1514     BlockDriverState *bs;
1515     IscsiLun *iscsilun = NULL;
1516     QDict *bs_options;
1517
1518     bs = bdrv_new("", &error_abort);
1519
1520     /* Read out options */
1521     total_size = DIV_ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
1522                               BDRV_SECTOR_SIZE);
1523     bs->opaque = g_new0(struct IscsiLun, 1);
1524     iscsilun = bs->opaque;
1525
1526     bs_options = qdict_new();
1527     qdict_put(bs_options, "filename", qstring_from_str(filename));
1528     ret = iscsi_open(bs, bs_options, 0, NULL);
1529     QDECREF(bs_options);
1530
1531     if (ret != 0) {
1532         goto out;
1533     }
1534     iscsi_detach_aio_context(bs);
1535     if (iscsilun->type != TYPE_DISK) {
1536         ret = -ENODEV;
1537         goto out;
1538     }
1539     if (bs->total_sectors < total_size) {
1540         ret = -ENOSPC;
1541         goto out;
1542     }
1543
1544     ret = 0;
1545 out:
1546     if (iscsilun->iscsi != NULL) {
1547         iscsi_destroy_context(iscsilun->iscsi);
1548     }
1549     g_free(bs->opaque);
1550     bs->opaque = NULL;
1551     bdrv_unref(bs);
1552     return ret;
1553 }
1554
1555 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1556 {
1557     IscsiLun *iscsilun = bs->opaque;
1558     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1559     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1560     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1561     return 0;
1562 }
1563
1564 static QemuOptsList iscsi_create_opts = {
1565     .name = "iscsi-create-opts",
1566     .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
1567     .desc = {
1568         {
1569             .name = BLOCK_OPT_SIZE,
1570             .type = QEMU_OPT_SIZE,
1571             .help = "Virtual disk size"
1572         },
1573         { /* end of list */ }
1574     }
1575 };
1576
1577 static BlockDriver bdrv_iscsi = {
1578     .format_name     = "iscsi",
1579     .protocol_name   = "iscsi",
1580
1581     .instance_size   = sizeof(IscsiLun),
1582     .bdrv_needs_filename = true,
1583     .bdrv_file_open  = iscsi_open,
1584     .bdrv_close      = iscsi_close,
1585     .bdrv_create     = iscsi_create,
1586     .create_opts     = &iscsi_create_opts,
1587     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1588
1589     .bdrv_getlength  = iscsi_getlength,
1590     .bdrv_get_info   = iscsi_get_info,
1591     .bdrv_truncate   = iscsi_truncate,
1592     .bdrv_refresh_limits = iscsi_refresh_limits,
1593
1594     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1595     .bdrv_co_discard      = iscsi_co_discard,
1596     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1597     .bdrv_co_readv         = iscsi_co_readv,
1598     .bdrv_co_writev        = iscsi_co_writev,
1599     .bdrv_co_flush_to_disk = iscsi_co_flush,
1600
1601 #ifdef __linux__
1602     .bdrv_ioctl       = iscsi_ioctl,
1603     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1604 #endif
1605
1606     .bdrv_detach_aio_context = iscsi_detach_aio_context,
1607     .bdrv_attach_aio_context = iscsi_attach_aio_context,
1608 };
1609
1610 static QemuOptsList qemu_iscsi_opts = {
1611     .name = "iscsi",
1612     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1613     .desc = {
1614         {
1615             .name = "user",
1616             .type = QEMU_OPT_STRING,
1617             .help = "username for CHAP authentication to target",
1618         },{
1619             .name = "password",
1620             .type = QEMU_OPT_STRING,
1621             .help = "password for CHAP authentication to target",
1622         },{
1623             .name = "header-digest",
1624             .type = QEMU_OPT_STRING,
1625             .help = "HeaderDigest setting. "
1626                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1627         },{
1628             .name = "initiator-name",
1629             .type = QEMU_OPT_STRING,
1630             .help = "Initiator iqn name to use when connecting",
1631         },
1632         { /* end of list */ }
1633     },
1634 };
1635
1636 static void iscsi_block_init(void)
1637 {
1638     bdrv_register(&bdrv_iscsi);
1639     qemu_add_opts(&qemu_iscsi_opts);
1640 }
1641
1642 block_init(iscsi_block_init);